No Priors: Artificial Intelligence | Technology | Startups
00:17:52 · Elad Gil, Sarah Guo
[00:17:52] S0: Perceived scale. Perceived Yeah. Perceived scale is the important angle. So the other thing a lot of people out here are working on or talking about is if you talk to people at the labs, there's this enormous manic energy right now. We're six months ish or towards the end of the year to be completely done with code, like, as a solved problem. And then we'll probably hit some form of light RSI by end of next year. And at that point, you'll have models training big chunks of the models themselves. I think it's probably more post training initially, maybe it could impact pre training over time more quickly as well. And because of that, many people believe, hey, if I have a year, year and a half left of productive work in my career, I should be working sixteen hours a day because every week is, you know, 2% of all the time I have left to be productive before I get displaced by AI. What do you think of that? [00:18:45] S1: Like, do I believe it, or what happens if it's true? [00:18:48] S0: Do you believe it? [00:18:49] S1: I think the idea that the models can improve their own training if the leading scientists working on this believe it, and it's an extension of what we are already seeing in code and math, of course, you should believe it. Right? [00:19:07] S0: Yeah. But on that timeline.
00:19:07 · Elad Gil, Sarah Guo
[00:19:07] S0: Yeah. But on that timeline. [00:19:08] S1: The data I have is that a number of very smart and even very self aware research scientists have felt that, you know, there was some knee in the curve on recursive self improvement or ASI eighteen months away, every eighteen months for the last five years. So how good of it is a predictor? It's not clear. I think the, like, the extension from code to training code to data pipeline work, is way easier to believe. The question of like how you are going to go gather that data for less verifiable, more complex domains? Or do you run into the actual constraints on the, like, physical compute accessibility side? Like, I think that's probably more of a limiter than, this being algorithmically possible.
00:20:00 · Elad Gil, Sarah Guo
[00:20:00] S0: Yeah. I mean, the physical compute basically reinforces an oligopoly market because what it does is it creates a ceiling on the rate of progress any single lab can get if effectively assume the compute is roughly pro rata across the ecosystem to the big labs. And so in the absence of a lack of compute constraints, you're almost having an enforced oligopoly market up to a point, or at least you force closer competition between the players than what exists otherwise, which I think is an interesting artifact of this moment in time. And the question is when does that lift and what does that look like? So, yeah, it's kind of it's a very exciting time. I mean, I always wonder about second order effects of that belief of it's eighteen months away because that does suggest there could be a burnout cycle in eighteen months. Like, I know some people at one of the major labs who for a while, brought up with me, should they get married? Do people want to get married? Should they get married? Because I don't know what happens in eighteen months to the world. It's like, you should get married. You should go ahead. It'll be okay. And so I do think we're living through this very manic, very exciting, very intense work period. So, yeah, it's really fun stuff. [00:21:01] S1: I think it's kind of tragic, man. [00:21:03] S0: Really? Why?
00:21:03 · Elad Gil, Sarah Guo
[00:21:03] S0: Really? Why? [00:21:04] S1: I think the reactions of some really extraordinary research friends to it, it feels a little bit tragic. I think it like I feel like it's psychologically most similar to if people think they're gonna die. Right? Like, how would you spend the last two years of your life? Would you spend it the way you are today, or would you spend it in a very different way is a, like, a not unrelated philosophical question. And so you do have people who are like, ah, like, my contribution is a bit irrelevant given RSI in the next eighteen months. So should I get married? Should I bother to work? Should I travel? Should I only work? And, you know, I just actually think it's a much more stable and satisfying state if people act as if they have. But maybe you think that's blind.
00:21:49 · Elad Gil
[00:21:49] S0: Yeah. I just think there's a lot of second order effects that are happening or going to happen. Part of them are driven by this belief system and potential burnout over time. Part of it is going to be One thing that I've noticed that is happening at some of the labs is that as compute becomes really the scarce resource, it turns out that there's say a few dozen researchers that drive 80% of the results at any given place, which is a really interesting human power law. If you actually look at it in any field, there's at most a few dozen people who drive the field. You look at breast cancer research, you look at certain subfields of mathematics, you look at subfields of physics, You look at the entrepreneurial ecosystem and founders. There's a handful of people, dozens of people who drive most progress. And that also happens in AI research. And increasingly, compute is differentially provided to those people. Right? And so I know some labs have slowed down on their hiring of researchers unless they're above a very, very high bar because the cost isn't the researcher, it's the compute associated with the person. That's where the real bottleneck is. I think there's this broader concept of return on invested tokens, like an ROI kind of metric, which is if you have a certain token budget, who do you give it to and why? This is kind of like engineering back in the day. Right? The internal tools teams at companies were always starved for resources because many, at least tech companies, would rather use the same engineers to build product than to build internal tools that would make other functions more productive. That's why I think the death of SaaS is a little bit overstated because why would you use tokens on a bunch of SaaS stuff that you're not actually paying that much for per year relative to the outcome of those same tokens being invested against a core product or against some massive margin lift or some other thing? And so I think increasingly we've shifted from a world where people said, Hey, everybody use AI and do whatever you want, to hey, we have to measure, spend, and move more things to open source. And then I think the next wave is what are the projects and people that should actually get outsized pieces of a token budget, and what is that return on investment? It's this sort of next shift that's coming? It'll take some time, though.
00:23:57 · Elad Gil, Sarah Guo
I mean, many people are still at like, hey, everybody try AI or whatever, you know, big enterprises. [00:24:02] S1: What do you think is the appropriate compute, like token budget for a business or a human being three to five years from now? Should I look at it like rent?